Fiction Eval

Which AI writes the best fiction?

We put the top AI models head-to-head on real genre scenes — LitRPG, cozy mystery, romance, thriller, and more — and had them judged blind, by a different AI than the ones being graded. Here's who actually writes the best fiction. Updated whenever a new model drops.

loading…

How to read it: higher score = better writing. When two models are close enough that we can't reliably tell them apart, they're marked as tied — a tiny gap isn't a real difference. Slop is a separate measure of AI tells (lower is cleaner).

Updated July 2026 — re-run within days of each major model release.

Score overall writing strength, like a chess rating, with its likely range below · Record wins–losses–ties · Slop AI-tell density per 1,000 words (lower is cleaner) · Voice and Cleanness the judge's 1–5 ratings for a distinct voice and freedom from AI tells. Full method and every prompt are further down — or read the report or grab the raw data (JSON).

Play judge yourself

Two passages from the same scene, names hidden. Pick the one you'd rather read — then we reveal who wrote each, and whether you agreed with the board.

loading…

Does your own writing read as AI?

Paste a scene or chapter (120+ words). It gets scored against the same checklist the board uses — the "it wasn't X, it was Y" crutch, essay-ish vocabulary, stock phrases, filler words. It's a fixed checklist, not another AI's opinion, and nothing you paste ever leaves your browser.

Read the actual writing

The passages are the whole point — the scores come from these. Pick a scene:

FAQ

Which AI model writes the best fiction?

As of the July 2026 board, Claude Opus 4.8 and Claude Sonnet 4.6 are tied for first — they're close enough that it's a coin flip, and run it again either could come out on top. Grok 4.3 is a clear third and GLM 5.2 fourth. The rankings come from 819 blind head-to-head matchups across 10 genre scenes, three tries per model. When two models' score ranges overlap, treat them as tied rather than reading anything into a small gap.

How is the leaderboard judged?

Every pair of models is judged blind on the same scene, and we run it twice with the two passages in swapped order so being shown first can't help anyone. The judge is GPT-5.4 at temperature 0 — deliberately a model from a family that isn't on the board, so no contestant ever scores its own side. The models write at temperature 0.9 (normal creative settings), three tries per scene. We tally all the wins and losses into a single strength score (shown like a chess rating) and re-run the numbers many times over to get the "likely range" next to each one. When two ranges overlap, we call it a tie instead of faking a winner. All ten prompts and every generated passage are shown on this page.

What is AI slop and how do you measure it?

"Slop" is the stuff that makes writing read machine-made: the "it wasn't X, it was Y" crutch, essay words (delve, tapestry, testament), stock phrases ("a wave of dread washed over her"), hedging, and filler. We score it with a fixed checklist — no AI judgment, and the same passage always gets the same number — reported per 1,000 words. We keep it separate from the quality ranking on purpose, because clean prose and good prose aren't the same thing.

Can I check my own writing for AI tells?

Yes — the checker above scores any text you paste with the same checklist the board uses, and compares your number against all eight ranked models. It runs entirely in your browser; nothing you paste is uploaded anywhere.

How often does the leaderboard update?

The full evaluation re-runs within days of any major model release — new prompts are written fresh each cycle so they can't appear in training data. The email list below gets each new board when it lands.

Who makes this, and can I trust it?

It's built by a working genre author who drafts with these models daily, and it lives on The Book Factory (an AI writing-tools site) because that's who cares about the answer. Rather than ask you to take the ranking on faith, the whole thing is built to be checked: the judge is a model from a family that isn't ranked, so it can't crown itself; every prompt and every passage is public, so you can read them and disagree; and you can play judge blind above and see whether your own picks match the board. If a result looks wrong, the receipts are right here.

Deep dives

The July 2026 Fiction Model Report · every head-to-head comparison · genre boards: LitRPG, cozy mystery, romance, thriller, progression fantasy, xianxia, science fiction, fantasy

New models drop constantly. We re-run the whole board within days — get the updated rankings by email, nothing else, ever.